Back to Reframe Venture’s AI Homepage

A scenario analysis
2026 – 2035

Scenario I:
AI, Energy & Water

AI is straining power grids and watersheds even as it promises to accelerate the net-zero transition. Sometime around 2027, the paths diverge. Which fork we take is not inevitable. It will be chosen by voters and by capital allocators.

This article was written following an in-depth workshop in London bringing together leading academics, asset owners and managers, and venture capital investors in October 2025. It combines the insight of this group with an extensive review of the literature on the topic.

The essay version of this scenario was written by Tiffany Tsoi and has been reviewed by Prof. Charlie Wilson (University of Oxford), Dr. Gina Neff (University of Cambridge), Dr. Natasha McCarthy (Royal Academy of Engineering), Sophie Walker (EQT), Dr. Johannes Lenhard (Reframe Venture).

The workshop and this scenario are part of a series exploring AI’s systemic impacts in pursuit of alternative narratives. The others can be viewed here.

Supported by

Mid 2026

The Build-Out

The data-centre construction boom is in full swing. Hyperscalers, valuing the prospect of AGI in the hundreds of trillions of dollars, operate on a winner-takes-most logic: whoever scales compute fastest wins everything. The priority is speed-to-power — getting megawatts of compute power online quickly — not sustainability or efficiency.4

Rows of electrical switchgear cabinets inside a data-centre hall
Speed-to-power: hyperscale halls are provisioned faster than the grid can green them.

The economics are unforgiving. IT equipment turns obsolete in four to six years, so the risk of overbuilding is acute, destroying private capital and leaving society with significant physical waste if the demand curve disappoints.12

And the demand curve might disappoint. Multiple observers argue the headline projections are exaggerated.34 Even the CEO of Constellation Energy, one of the US’ largest independent power producers, cautioned investors on a 2025 earnings call that the industry has “a history of over forecasting” demand.5 The IEA’s own outlook spans a wide sensitivity band; an indication of how little anyone knows.6

Line chart of global data-centre electricity consumption 2020 to 2035 under IEA sensitivity cases: Lift-Off, Base, High Efficiency, and Headwinds, diverging widely after 2024
Global data-centre electricity consumption by sensitivity case, 2020–2035 (TWh). Figure reproduced from the original Reframe Venture scenario, after the IEA.6
Expand: why the forecasts may be inflated

Data-centre developers file duplicate grid-connection requests across multiple utilities while site-shopping, so interconnection queues may overstate real demand. Institutional analyses from IEEFA, Man Group, and McKinsey each flag versions of the same risk: a multi-trillion-dollar race to scale compute, priced as if every projected megawatt materialises.134 If it doesn’t, the overhang lands on utilities, ratepayers — and the investors who financed the build.

Meanwhile the ecosystem is fragmented. Founders, big tech, energy experts, regulators, and investors each assume someone else needs to act. What little change is happening is driven primarily by customers, particularly through enterprise procurement processes, acting as a critical ‘signalling effect’, which investors can then use to mobilise action.

Late 2026

The Grid Pushes Back

The physical world begins to answer. Because AI’s anticipated growth curve is steeper than the grid’s, clean power, transmission capacity, and transformers are diverted to data centres — potentially slowing decarbonisation everywhere else.7 The Netherlands, Singapore, and other jurisdictions maintain de facto moratoria on new data centres, their grids too congested to take more.8

The marginal electron is not clean: BloombergNEF projects roughly 64% of incremental data-centre power will come from fossil fuels, with the lifespans of existing coal and gas plants quietly extended.9 Capital earmarked for clean energy is redirected — the UAE’s primary renewable energy firm, Masdar, recently shifted billions of dollars from planned green hydrogen projects to artificial intelligence and data centre infrastructure.10

Locally, the strain is already political. Consumer electricity bills are rising notably in Maryland, Virginia, and Oregon. Residents are pushing back and officials feel the pressure, while US utilities scramble to make tech firms pay their fair share of infrastructure costs.11

Expand: the water problem

Training a large AI model can consume millions of litres of water, and discharge carries contamination risks. Water use varies enormously by location and time of day — which also means siting and scheduling choices can dramatically reduce it. Communities near natural water sources are already organising against data-centre proposals.12

Transparency norms are forming, though. The EU AI Act mandates energy-efficiency disclosure for models, providing initial transparency and allowing customers to compare different providers. Salesforce — with Hugging Face, Cohere, and Carnegie Mellon — releases an AI Energy Score benchmarking models on a 1–5 scale, and commits to disclosing energy data for its own proprietary models.13 The measurement infrastructure for a different path now exists. Whether anyone uses it is the open question.

Mid 2027 · Branch point

The Fork in the Road

By mid-2027, the contradiction can no longer be deferred. Grid queues, water permits, procurement standards, and LP pressure all force the same question: does the AI build-out internalise its resource footprint, or not?

Nothing about what follows is inevitable. The decision tree is still broad, and it is shaped at critical junctures by innovators and the people who fund them. Two trajectories dominate the scenario space.

Load Growth

Late 2027

The Race Continues

Continued focus on AGI-at-all-costs, in combination with geopolitical competition, settles the argument in favour of ever-larger, more complex models. There is little momentum for greening the AI stack, and data-centre energy demand grows unabated. Speed-to-power wins every siting decision, and the fossil share of the marginal grid mix hardens.9

Black-and-white photograph of power-station smokestacks releasing plumes over a treeline and lake
Life extension: coal and gas plants stay online to feed short-term data-centre demand.

2028

Backlash Without Teeth

Consumer electricity bills skyrocket in data-centre regions.11 Discharge and contamination of potable water worsens, droughts arrive more often and hit harder. Local water-quality indicators — temperature, pH, total dissolved solids, conductivity — plummet near clusters of compute.12

Sociopolitical backlash foments. Critical media coverage moves from the business pages to the front page, and sentiment analysis of headlines trends steadily negative. Legislative and regulatory activity escalates — hearings, draft bills, town halls — but fails, again and again, to materialise into binding law.

2029

Two Ways Down

From here the negative trajectory itself forks, depending on whether AI adoption succeeds commercially. Both sub-paths end badly for the climate — through different mechanisms.

Sub-scenario · Adoption succeeds

The Rebound

AI developers crack value realisation at scale. Trust is prioritised, enterprises and consumers adopt en masse, and energy-efficiency bottlenecks are engineered away. And it doesn’t help: every efficiency gain is swamped by the explosion in total consumption — a rebound effect. The climate problem worsens despite optimisation.

The second-order effects are worse. Mass job displacement sends inequality and social polarisation to unprecedented levels, and as a result, the social consensus frays and democratic institutions weaken. The political coalition behind climate action erodes just when it is needed most.14

Sub-scenario · Adoption stalls

The Stall

Alternatively, value realisation never arrives. Energy-supply bottlenecks persist, trustworthiness deficits remain, and adoption falters. The public — deeply distrustful of AI systems, especially with personal data — never embraces consumer-facing climate-beneficial applications, which fail to scale.15

Eventually, the physical overbuild becomes undeniable. Completed data centres run at low utilisation with no clear repurposing; in-flight projects halt; gas infrastructure built for the boom is stranded — after having diverted capital from renewables.12 Hyperscaler, AI, and chip valuations correct drastically, effectively overnight. Private capital and credit vehicles absorb persistent, devastating write-downs. Construction and manufacturing follow. Populism rises and the economic crisis dominates discourse, diverting attention and capital further away from climate change.

Trading screen showing a steeply falling candlestick chart
The correction: valuations reprice overnight; private credit absorbs the write-downs.
Expand: what is a rebound effect?

When efficiency makes something cheaper, we consume more of it — sometimes so much more that total resource use rises. A model that is 10× more efficient but deployed 1,000× more widely is a net increase. In this branch, efficiency engineering succeeds while efficiency outcomes fail, because nothing constrains aggregate demand.

2032 → 2035

The Hot Decade

Frontier climate innovation, underfunded through the crash years, fails at scale. The efficiency gains AI did deliver show up more in oil-and-gas prospecting than in climate solutions. Fossil plants brought online for short-term data-centre demand keep running.

Global CO₂ emissions rise beyond previous anticipations. Temperatures climb, greenhouse-gas concentrations increase, and extreme weather grows in frequency and intensity. Food security declines and populations are displaced, with profound societal and economic consequences. The decade AI was supposed to buy back was spent instead.

Efficiency Push

Late 2027

Pricing the Externality

In this branch, the pressure points bite. The largest energy- and water-intensive models face regulatory scrutiny, reputational risk, and — decisively — procurement barriers: enterprise B2B customers, wanting lower compute bills anyway, start requiring sustainability metrics in contracts. ‘Greener’ becomes ‘cheaper’ becomes ‘table stakes’.

Capital allocation shifts up the chain. Asset owners press fund managers with specific, non-negotiable environmental governance questions. VCs, in turn, push founders toward a “full system approach” from the earliest stages of development.

Wind turbine at dusk in front of desert mountains
Procurement pressure moves the marginal megawatt toward clean supply.

2028

Measurement Becomes Market Practice

The AI Energy Score matures from a benchmark into a de facto procurement standard, with major labs disclosing under the EU AI Act’s efficiency provisions.13 Asset owners go beyond screening: they create dedicated funds and allocations that shift VC incentives and de-risk climate-positive AI investment, balancing their own climate exposure with climate-technology upside.

Money flows into energy- and water-efficient AI infrastructure — clean-energy data centres, waste-heat reuse feeding district heating networks that displace fossil boilers in surrounding neighbourhoods.21 Momentum builds for climate-positive AI use-cases, because for the first time they are cheaper to fund than to ignore.

2029 → 2031

The Application Wave

With the stack greening, attention turns to what AI can do for the transition. Climate-positive AI startups attract high valuations. AI accelerates discovery in catalysts, novel materials, and battery chemistries.16

Hands swapping a green battery pack into an electric cargo bike
Discovery compounds: AI-accelerated battery chemistry reaches deployment.

Small, specialised models — trained on high-quality, localised or proprietary data in pharmaceuticals, finance, and agricultural measurement — outperform giant generalists on the tasks that matter, while aligning with sustainability and data-protection norms.1819 AI-assisted grid management handles intermittent renewables, onboards data centres faster, and runs them flexibly; training jobs are dynamically scheduled around real-time local weather, and inference is routed to where water is abundant.1220

Quietly transformative: AI translation and language unification aggregate fractured research across science and policy, accelerating R&D and enabling better modelling of complex futures — and better policymaking.

2032 → 2035

The Lever

By the early 2030s, AI’s pivotal role in reaching net zero is being realised rather than merely projected. The applications compound: optimised fuel mixes in cement kilns; route selection and EV adoption cutting vehicle emissions; HVAC control across building complexes; methane-leak detection over oil and gas operations; renewables-heavy grids kept stable and efficient; even protein-property identification improving meat alternatives.22

Data-centre demand still grows — but on a flatter curve, powered cleaner, sited smarter, and repaid many times over in avoided emissions elsewhere. The externality was priced, so the technology could be spent where it counted.

Either way

Takeaways for Investors

The central thesis of the scenario: AI’s future climate impact is not inevitable. The decision tree remains broad at critical junctures, and it is shaped by innovators and capital allocators. Three areas of the “greening the stack” opportunity remain underfunded:

Reducing the cost of AI

Software optimisation cuts energy use with substantial cost savings and minimal capability loss.17 Small Language Models need far less compute to train and run, with impressive capability from well-designed architectures — plus better privacy, reduced bias, and relevance in data-scarce emerging markets.1819

Grid-management efficiency

AI-assisted handling of complex electricity networks: managing intermittent renewables, onboarding data centres faster, operating them flexibly, scheduling training around local weather, and siting inference for water efficiency.1220

Greener data centres

Flexibility, cleaner energy and fuel sources, and waste-heat reuse — district heating that eliminates the need for fossil-fuel heating in surrounding neighbourhoods.21

Six practical actions

  1. Measure the portfolio. Collect and monitor energy consumption, water usage, and carbon intensity across AI holdings; benchmark against industry standards.
  2. Push the full-system approach early. Energy mix, water impact, and environmental trade-offs against model performance, from the earliest stages.
  3. Stress-test the infrastructure. Assess data-centre exposure to heat stress, water scarcity, and extreme weather.
  4. Fund right-sized models. Support SLM development and testing platforms (e.g. Hugging Face’s HuggingChat).
  5. Watch the sentiment. Monitor public opinion and media coverage of data-centre build-out in key regions — a leading indicator of regulatory and reputational risk.
  6. Know your own systemic footprint. Understand your organisation’s aptitude for contributing to — or mitigating — systemic risk.

References

Sources

All sources below are retained from the original Reframe Venture analysis.

  1. IEEFA (2025), The risk of AI-driven overbuilt infrastructure is real.
  2. Lovins (2025), Data centers, AI, and demand response (Stanford, PDF).
  3. Man Group (2026), The AI bubble.
  4. McKinsey (2025), The cost of compute: a $7 trillion race to scale data centers.
  5. Constellation Energy Q1 2025 earnings-call transcript (Joseph Dominguez remarks).
  6. IEA (2025), Global data centre electricity consumption by sensitivity case, 2020–2035.
  7. Hiller & Ramkumar (2024), Big tech is rushing to find clean power to fuel AI’s insatiable appetite, WSJ.
  8. Wilson, Fan & Amanta (2025), Oxford Energy Forum: AI & energy (PDF).
  9. BloombergNEF, New Energy Outlook.
  10. Ratcliffe & Paola (2025), UAE’s renewables champion shifts hydrogen billions to data boom, Bloomberg.
  11. Saul et al. (2025), AI data centers and electricity prices, Bloomberg Graphics.
  12. Li et al. (2025), Making AI less thirsty, ACM.
  13. Salesforce (2025), AI Energy Score (with Hugging Face, Cohere, Carnegie Mellon).
  14. Amanta & Wilson (2025), The overlooked climate risks of artificial intelligence, Oxford Energy Day.
  15. Vrain & Wilson (2025), Data concerns undermine AI applications for climate.
  16. Wilson, Fan & Amanta (2025), Oxford Energy Forum 145 (PDF).
  17. Frugal AI Hub (2026), frugalai.org.
  18. Belcak et al. (2025), Small language models are the future of agentic AI, arXiv.
  19. Nanni et al. (2025), Why we still need small language models, even in the age of frontier AI, Alan Turing Institute.
  20. IEA (2025), Energy and AI.
  21. World Economic Forum (2025), 6 ways data centres can cut emissions.
  22. Stern et al. (2025), AI and climate, npj Climate Action.
  23. Reframe Venture (2026), Scenario I: Energy / Water — the original analysis this timeline adapts.